Inference
EN

Gemini 3.5 Flash vs Gemma 4 12B

Change models
Gemini 3.5 FlashGoogle
Gemma 4 12BGoogle
IndexAll categories, out of 100.
64.2 ±2.8 best
39.0 ±4.0
GapIndex points below the top.
Top
−25.2
CoverageIndex weight that has results.
95%
100% best
SpeedOutput tokens per second.
137/s
—
EloLMArena rating and rank.
1483
—
ReleasedMonth of release.
2026-05
2026-06
Input / 1MUS dollars per 1M input tokens.
$1.50
—
Output / 1MUS dollars per 1M output tokens.
$9.00
—
Batch output / 1MBatch price per 1M output tokens.
$4.50
—
1K chats2K tokens in, 500 out, each.
$7.50
—
1K documents20K tokens in, 1K out, each.
$39.00
—
1K agent tasks200K tokens in, 20K out, each.
$480.00
—
ContextMaximum tokens in one request.
1.05M best
256K
AgenticMulti-step tasks with tools.
64.7 best
30.5
CodingCode writing and repair.
60.8 best
43.5
ReasoningLogic problems and puzzles.
62.2 best
41.5
MultimodalTasks with images and text.
64.7 best
36.2
KnowledgeFacts and expert knowledge.
68.5 best
41.5
MultilingualTasks in many languages.
—
29.7
InstructionTasks with strict prompt rules.
69.9 best
57.7
MathMath problems.
61.5 best
41.1

Benchmarks

13 shared
Only Gemini 3.5 Flash: 78Only Gemma 4 12B: 15
Gemini 3.5 Flash
Gemma 4 12B
Artificial Analysis Humanity's Last ExamKnowledge
42.7% best36 / 222
15.7%126 / 222
Artificial Analysis Long Context ReasoningReasoning
69.3% best115 / 217
63.7%137 / 217
Critical Physics TasksReasoning
13.1% best55 / 214
0.0%150 / 214
Artificial Analysis GPQA DiamondKnowledge
92.2% best21 / 199
75.3%115 / 199
τ²-Bench Tool-Agent-User EvaluationAgentic
95.3% best21 / 149
36.3%108 / 149
Artificial Analysis IFBenchInstruction
76.3% best13 / 146
73.5%28 / 146
Graduate-Level Google-Proof Q&AKnowledge
92.2% best20 / 141
78.8%71 / 141
GDPval-AA normalizedAgentic
42.2% best45 / 133
0.0%101 / 133
Artificial Analysis MMMU-ProMultimodal
84.3% best8 / 113
69.7%76 / 113
GPQA Diamond · BenchLMKnowledge
92.7% best16 / 84
78.8%64 / 84
Massive Multi-discipline Multimodal Understanding ProMultimodal
83.6% best4 / 53
69.1%45 / 53
ExtractBench: schema-guided enterprise document extraction (LlamaIndex leaderboard)Multimodal
79.8% best21 / 26
45.4%26 / 26
MRCRv2Reasoning
77.3% best6 / 14
43.4%12 / 14